Deadlock-Resistant Flow Control Procedures for Store-and-Forward Networks

A flow control procedure for an acyclic store-and-forward network is introduced which uses only information local to each node and is deadlock-resistant (i.e., detects and recovers deadlock at a negligible cost). The procedure requires less than buffers (the exact value depending on the network topology), where is the number of nodes in the network. It is shown that this number is lower bound for distributed deadlock-resistant procedures. From a practical point of view, this means that a single buffer class is sufficient, provided that the exchange mechanism for input buffers is realized as a synchronized procedure between any two contiguous nodes (rather than as a purely local procedure). Some extension to more general networks are then proposed.